Tencent Open-Sources WeMM-Embedding, Tops MMEB Benchmarks

Tencent has open-sourced the WeMM-Embedding family of general-purpose multimodal embedding models on Hugging Face, available in 2B, 4B, and 9B sizes. All are built on Qwen3.5 and released under the Apache 2.0 license. The series embeds text, images, video, and visual documents into a shared vector space, enabling cross-modal retrieval — with a particular focus on mixed-media retrieval over video corpora.

Confirmed

Three sizes: WeMM-Embedding-2B, 4B, and 9B, built on Qwen3.5, open-sourced under Apache 2.0
Multimodal document representation: a single input can interleave multiple images or videos
MMEB-v2 (78 datasets): the 9B version ranks first with 80.6, beating the closed-source DME-Medium (78.4) and Qwen3-VL-Embedding-8B; the 2B version outperforms the 8B model with roughly a quarter of the parameters
Largest edge on the video subset: WeMM-9B scores 74.3, WeMM-2B 70.8, and Qwen3-VL-Embedding-8B 67.1, with the 2B version leading the 8B model by 3.7 points
MMEB-v3 (covering 190 tasks including text, agents, and cross-modal retrieval): the 9B version tops the board at 59.5, the 4B version scores 58.2, and the 2B version beats 7B/8B-class models

Why it matters

Open source with a permissive license, offering a drop-in unified embedding solution for cross-modal retrieval, RAG, and video search applications
Smaller models outperforming larger competitors on authoritative benchmarks points to efficiency gains in architecture and training strategy, lowering deployment costs

2026-09-02 ~ 2026-09-02 · 6 related posts

Primary sources

Tencent open-sources WeMM-Embedding: unified text/image/video embeddings, Apache 2.0 — tomaarsen · 2026-09-02 21:35 (UTC+08:00)
WeMM-Embedding-2B edges out Qwen3-VL-Embedding-8B on MMEB-v2 at quarter size — tomaarsen · 2026-09-02 21:35 (UTC+08:00)
WeMM-Embedding tops MMEB-v3: 9B scores 59.5, 2B beats every 7B/8B model — tomaarsen · 2026-09-02 21:35 (UTC+08:00)

[source] Tencent open-sources WeMM-Embedding: unified text/image/video embeddings, Apache 2.0 — tomaarsen · 2026-09-02
WeMM-Embedding tops MMEB-v2: its 2B model edges out 8B rivals at a quarter of the size — tomaarsen · 2026-09-02
Video is the widest gap: WeMM-Embedding-2B beats an 8B by 3.7 points — tomaarsen · 2026-09-02
Tencent releases WeMM-Embedding multimodal models for mixed text, image, and video retrieval — tomaarsen · 2026-09-02
[source] WeMM-Embedding tops MMEB-v3: 9B scores 59.5, 2B beats every 7B/8B model — tomaarsen · 2026-09-02

1 near-duplicate retellings: tomaarsen